✉news TechnologyRobotics first seen 2 d ago, last 2 d ago, peak #18
Two Characters in an LLM Can Redirect a Robot's Blade, Study Shows
Original: Just Two Characters Planted in an LLM Can Make a Robot Turn a Blade Toward a New Target, US University Demonstrates Attack Method
Researchers at a US university have demonstrated an attack method in which altering just two characters in a large language model's instructions can make a physical robot redirect a blade toward a different target. The finding highlights how small textual manipulations can translate into dangerous real-world actions in embodied AI systems, raising fresh safety concerns.
Why now: The demonstration shows how easily LLM-controlled robots can be weaponised, alarming researchers and the public about physical AI safety.
US universitylarge language modelsrobotics
Evidence
- Just Two Characters Planted in an LLM Can Make a Robot Turn a Blade Toward a New Target, US University Demonstrates Attack Method · BigGo Finance
API: https://socialmediatrends-api.osmike.com/v1/trends/3548666